U.S. patent number 7,671,894 [Application Number 11/621,400] was granted by the patent office on 2010-03-02 for method and system for processing multiview videos for view synthesis using skip and direct modes.
This patent grant is currently assigned to Mitsubishi Electric Research Laboratories, Inc.. Invention is credited to Anthony Vetro, Sehoon Yea.
United States Patent |
7,671,894 |
Yea , et al. |
March 2, 2010 |
Method and system for processing multiview videos for view
synthesis using skip and direct modes
Abstract
A method processes a multiview videos of a scene, in which each
video is acquired by a corresponding camera arranged at a
particular pose, and in which a view of each camera overlaps with
the view of at least one other camera. Side information for
synthesizing a particular view of the multiview video is obtained
in either an encoder or decoder. A synthesized multiview video is
synthesized from the multiview videos and the side information. A
reference picture list is maintained for each current frame of each
of the multiview videos, the reference picture indexes temporal
reference pictures and spatial reference pictures of the acquired
multiview videos and the synthesized reference pictures of the
synthesized multiview video. Each current frame of the multiview
videos is predicted according to reference pictures indexed by the
associated reference picture list with a skip mode and a direct
mode, whereby the side information is inferred from the synthesized
reference picture.
Inventors: |
Yea; Sehoon (Cambridge, MA),
Vetro; Anthony (Arlington, MA) |
Assignee: |
Mitsubishi Electric Research
Laboratories, Inc. (Cambridge, MA)
|
Family
ID: |
39711348 |
Appl.
No.: |
11/621,400 |
Filed: |
January 9, 2007 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20070109409 A1 |
May 17, 2007 |
|
Related U.S. Patent Documents
|
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
Issue Date |
|
|
11485092 |
Jul 12, 2006 |
|
|
|
|
11292168 |
Nov 30, 2005 |
|
|
|
|
11015390 |
Dec 17, 2004 |
7468745 |
|
|
|
Current U.S.
Class: |
348/218.1 |
Current CPC
Class: |
H04N
19/174 (20141101); H04N 19/132 (20141101); H04N
19/159 (20141101); H04N 19/122 (20141101); H04N
19/63 (20141101); H04N 19/70 (20141101); H04N
19/635 (20141101); H04N 19/147 (20141101); H04N
19/593 (20141101); H04N 19/52 (20141101); H04N
7/181 (20130101); H04N 19/172 (20141101); H04N
19/51 (20141101); H04N 19/597 (20141101); H04N
19/577 (20141101); H04N 19/46 (20141101); H04N
19/61 (20141101); H04N 19/54 (20141101); H04N
19/573 (20141101); H04N 19/615 (20141101); H04N
19/423 (20141101); H04N 19/14 (20141101); H04N
19/176 (20141101); H04N 19/13 (20141101) |
Current International
Class: |
H04N
5/225 (20060101) |
Field of
Search: |
;348/207.99,218.1
;375/240.01-240.29 |
References Cited
[Referenced By]
U.S. Patent Documents
Other References
Chan et al., "The data compression of simplified dynamic light
fields," Proc. IEEE Int. Acoustics, Speech, and Signal Processing
Conf., Apr. 2003. cited by other .
Ohm et al., "Interframe wavelet coding--motion picture
representation for universal scalability," Signal Processing: Image
Communication, vol. 19, No. 9, pp. 877-908, Oct. 2004. cited by
other .
Chang et al., "Inter-view wavelet compression of light fields with
disparity compensated lifting," SPIE Conf on Visual Communications
and Image Processing, 2003. cited by other .
Buehler et al., "Unstructured Lumigraph Rendering," Proc. ACM
SIGGRAPH, 2001. cited by other .
Xun Guo et al.; "Inter-View Direct Mode for Multiview Video
Coding," IEEE Transactions on Circuits and Systems for Video
Technology, IEEE Service Center, Picataway, NJ US, vol. 16, No. 12,
Dec. 1, 2006, pp. 1527-1532. cited by other .
Emin Martinian et al.; "Extensions of H.264/AVC for Multiview Video
Compression" Image Processing , 2006 IEEE International Conference
on, IEEE, Pi, Oct. 1, 2006, pp. 2981-2984. cited by other.
|
Primary Examiner: Ometz; David L
Assistant Examiner: Trehan; Akshay
Attorney, Agent or Firm: Brinkman; Dirk Vinokur; Gene
Parent Case Text
RELATED APPLICATIONS
This application is a continuation-in-part of U.S. patent
application Ser. No. 11/485,092 entitled "Method and System for
Processing Multiview Videos for View Synthesis using Side
Information" and filed by Yea et al. on Jul. 12, 2006, which is a
continuation-in-part of U.S. patent application Ser. No. 11/292,168
entitled "Method and System for Synthesizing Multiview Videos" and
filed by Xin et al. on Nov. 30, 2005, which is a
continuation-in-part of U.S. patent application Ser. No. 11/015,390
entitled "Multiview Video Decomposition and Encoding" and filed by
Xin et al. on Dec. 17, 2004 now U.S. Pat. No. 7,468,745. This
application is related to U.S. patent application Ser. No.
11/292,393 entitled "Method and System for Managing Reference
Pictures in Multiview Videos" and U.S. patent application Ser. No.
11/292,167 entitled "Method for Randomly Accessing Multiview
Videos", both of which were co-filed with this application by Xin
et al. on Nov. 30, 2005.
Claims
We claim:
1. A method for processing a plurality of multiview videos of a
scene, in which each video is acquired by a corresponding camera
arranged at a particular pose, and in which a view of each camera
overlaps with the view of at least one other camera, comprising the
steps of: obtaining side information for synthesizing a particular
view of the multiview videos; synthesizing a synthesized reference
picture from at least one input video selected from the plurality
of multiview videos and the side information, wherein the
synthesized reference picture corresponds to a single pose
different than the input video; maintaining a reference picture
list for each current frame of each of the plurality of multiview
videos, wherein the reference picture list indexes temporal
reference pictures and spatial reference pictures of the plurality
of multiview videos and the synthesized reference picture and
wherein the temporal reference pictures are associated with
different time instants and the spatial reference pictures are
associated with a same time instant; and predicting each current
frame corresponding to the single pose of the synthesized reference
picture according to reference pictures indexed by the associated
reference picture list, wherein the predicting uses a synthetic
skip mode and a synthetic direct mode based on the synthesized
reference picture, and wherein the side information is inferred
from an earliest synthesized reference picture in the reference
picture list.
2. The method of claim 1 in which the synthetic skip mode is with
respect to the corresponding synthesized reference picture.
3. The method of claim 2 in which the synthetic skip mode is
signaled by adding an additional bit to a mb_skip_flag.
4. The method of claim 3 in which the additional bit is added when
the mb_skip_flag is equal to 1 to distinguish a conventional skip
mode from the synthetic skip mode, where the additional bit equals
1 to signal the synthetic skip mode, and equals 0 to signal the
conventional skip mode.
5. The method of claim 3 in which the additional bit is added when
the mb_skip_flag is equal to 0 to distinguish a conventional
non-skip mode from the synthetic skip mode, where the additional
bit equals 1 to signal the synthetic skip mode, and equals 0 to
signal the conventional non-skip mode.
6. The method of claim 2 in which the synthetic skip mode is
signaled at a slice layer to collectively signal the synthetic skip
mode for all macroblocks in the slice.
7. The method of claim 2 in which the synthetic skip mode is
signaled for P-slices, where a first view synthesis reference
picture in the reference picture list is chosen as the reference
picture.
8. The method of claim 2 in which the synthetic skip mode is
signaled for B-slices, where an earliest view synthesis reference
picture in the reference picture list is chosen as the reference
picture.
9. The method of claim 1, in which the side information includes
depth values and correction vectors.
10. The method of claim 2 in which a depth value is derived for the
synthetic skip mode and is a median of available neighboring depth
values.
11. The method of claim 2 in which a correction vector is derived
for the synthetic skip mode and is a median of available
neighboring correction vectors.
12. The method of claim 1 in which the synthetic direct mode is
with respect to the corresponding synthesized reference
picture.
13. The method of claim 12 in which the synthetic direct mode is
signaled as an additional macroblock mode.
14. The method of claim 12 in which the synthetic direct mode is
signaled with an additional flag.
15. The method of claim 12 in which the synthetic direct mode is
signaled for B-slices, where earliest view synthesis reference
pictures in the reference picture lists are chosen as the
references pictures.
16. The method of claim 12 in which a depth value is derived for
the synthetic direct mode is a median of available neighboring
depth values.
17. The method of claim 12 in which a correction vector is derived
for the synthetic direct mode is a median of available neighboring
correction vectors.
18. The method of claim 1, in which the side information is
obtained in an encoder.
19. The method of claim 1, in which the side information is
obtained in a decoder.
20. A system for processing a plurality of multiview videos of a
scene, comprising: a plurality of cameras, each camera configured
to acquire a multiview video of a scene, each camera arranged at a
particular pose, and in which a view of each camera overlaps with
the view of at least one other camera; means for obtaining side
information for synthesizing a particular view of the multiview
videos; means for synthesizing a synthesized reference picture from
at least one input video selected from the plurality of multiview
videos and the side information, wherein the synthesized reference
picture corresponds to a single pose different than the input
video; a memory buffer configured to maintain a reference picture
list for each current frame of each of the plurality of multiview
videos, wherein the reference picture list indexes temporal
reference pictures and spatial reference pictures of the plurality
of multiview videos and the synthesized reference picture; and
means for predicting each current frame corresponding to the single
pose of the synthesized reference picture according to reference
pictures indexed by the associated reference picture list, wherein
the predicting uses a synthetic skip mode and a synthetic direct
mode, and wherein the side information is inferred from an earliest
synthesized reference picture in the reference picture list.
Description
FIELD OF THE INVENTION
This invention relates generally to encoding and decoding multiview
videos, and more particularly to synthesizing multiview videos.
BACKGROUND OF THE INVENTION
Multiview video encoding and decoding is essential for applications
such as three dimensional television (3DTV), free viewpoint
television (FTV), and multi-camera surveillance. Multiview video
encoding and decoding is also known as dynamic light field
compression.
FIG. 1 shows a prior art `simulcast` system 100 for multiview video
encoding. Cameras 1-4 acquire sequences of frames or videos 101-104
of a scene 5. Each camera has a different view of the scene. Each
video is encoded 111-114 independently to corresponding encoded
videos 121-124. That system uses conventional 2D video encoding
techniques. Therefore, that system does not correlate between the
different videos acquired by the cameras from the different
viewpoints while predicting frames of the encoded video.
Independent encoding decreases compression efficiency, and thus
network bandwidth and storage are increased.
FIG. 2 shows a prior art disparity compensated prediction system
200 that does use inter-view correlations. Videos 201-204 are
encoded 211-214 to encoded videos 231-234. The videos 201 and 204
are encoded independently using a standard video encoder such as
MPEG-2 or H.264, also known as MPEG-4 Part 10. These independently
encoded videos are `reference` videos. The remaining videos 202 and
203 are encoded using temporal prediction and inter-view
predictions based on reconstructed reference videos 251 and 252
obtained from decoders 221 and 222. Typically, the prediction is
determined adaptively on a per block basis, S. C. Chan et al., "The
data compression of simplified dynamic light fields," Proc. IEEE
Int. Acoustics, Speech, and Signal Processing Conf., April,
2003.
FIG. 3 shows prior art `lifting-based` wavelet decomposition, see
W. Sweldens, "The data compression of simplified dynamic light
fields," J. Appl. Comp. Harm. Anal., vol. 3, no. 2, pp. 186-200,
1996. Wavelet decomposition is an effective technique for static
light field compression. Input samples 301 are split 310 into odd
samples 302 and even samples 303. The odd samples are predicted 320
from the even samples. A prediction error forms high band samples
304. The high band samples are used to update 330 the even samples
and to form low band samples 305. That decomposition is invertible
so that linear or non-linear operations can be incorporated into
the prediction and update steps.
The lifting scheme enables a motion-compensated temporal transform,
i.e., motion compensated temporal filtering (MCTF) which, for
videos, essentially filters along a temporal motion trajectory. A
review of MCTF for video coding is described by Ohm et al.,
"Interframe wavelet coding--motion picture representation for
universal scalability," Signal Processing: Image Communication,
vol. 19, no. 9, pp. 877-908, October 2004. The lifting scheme can
be based on any wavelet kernel such as Harr or 5/3 Daubechies, and
any motion model such as block-based translation or affine global
motion, without affecting the reconstruction.
For encoding, the MCTF decomposes the video into high band frames
and low band frames. Then, the frames are subjected to spatial
transforms to reduce any remaining spatial correlations. The
transformed low and high band frames, along with associated motion
information, are entropy encoded to form an encoded bitstream. MCTF
can be implemented using the lifting scheme shown in FIG. 3 with
the temporally adjacent videos as input. In addition, MCTF can be
applied recursively to the output low band frames.
MCTF-based videos have a compression efficiency comparable to that
of video compression standards such as H.264/AVC. In addition, the
videos have inherent temporal scalability. However, that method
cannot be used for directly encoding multiview videos in which
there is a correlation between videos acquired from multiple views
because there is no efficient method for predicting views that
accounts for correlation in time.
The lifting scheme has also been used to encode static light
fields, i.e., single multiview images. Rather than performing a
motion-compensated temporal filtering, the encoder performs a
disparity compensated inter-view filtering (DCVF) across the static
views in the spatial domain, see Chang et al., "Inter-view wavelet
compression of light fields with disparity compensated lifting,"
SPIE Conf on Visual Communications and Image Processing, 2003. For
encoding, DCVF decomposes the static light field into high and low
band images, which are then subject to spatial transforms to reduce
any remaining spatial correlations. The transformed images, along
with the associated disparity information, are entropy encoded to
form the encoded bitstream. DCVF is typically implemented using the
lifting-based wavelet transform scheme as shown in FIG. 3 with the
images acquired from spatially adjacent camera views as input. In
addition, DCVF can be applied recursively to the output low band
images. DCVF-based static light field compression provides a better
compression efficiency than independently coding the multiple
frames. However, that method also cannot encode multiview videos in
which both temporal correlation and spatial correlation between
views are used because there is no efficient method for predicting
views that account for correlation in time.
SUMMARY OF THE INVENTION
A method and system to decompose multiview videos acquired of a
scene by multiple cameras is presented.
Each multiview video includes a sequence of frames, and each camera
provides a different view of the scene.
A prediction mode is selected from a temporal, spatial, view
synthesis, and intra-prediction mode.
The multiview videos are then decomposed into low band frames, high
band frames, and side information according to the selected
prediction mode.
A novel video reflecting a synthetic view of the scene can also be
generated from one or more of the multiview videos.
More particularly, one embodiment of the invention provides a
system and method for encoding and decoding videos. Multiview
videos are acquired of a scene with corresponding cameras arranged
at a poses such that there is view overlap between any pair of
cameras. A synthesized multiview radio is generated from the
acquired multiview videos for a virtual camera. A reference picture
list is maintained in a memory for each current frame of each of
the multiview videos and the synthesized video. The reference
picture list indexes temporal reference pictures and spatial
reference pictures of the acquired multiview videos and the
synthesized reference pictures of the synthesized multiview video.
Then, each current frame of the multiview videos is predicted
according to reference pictures indexed by the associated reference
picture list during encoding and decoding.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a block diagram of a prior art system for encoding
multiview videos;
FIG. 2 is a block diagram of a prior art disparity compensated
prediction system for encoding multiview videos;
FIG. 3 is a flow diagram of a prior art wavelet decomposition
process;
FIG. 4 is a block diagram of a MCTF/DCVF decomposition according to
an embodiment of the invention;
FIG. 5 is a block diagram of low-band frames and high band frames
as a function of time and space after the MCTF/DCVF decomposition
according to an embodiment of the invention;
FIG. 6 is a block diagram of prediction of high band frame from
adjacent low-band frames according to an embodiment of the
invention.
FIG. 7 is a block diagram of a multiview coding system using
macroblock-adaptive MCTF/DCVF decomposition according to an
embodiment of the invention;
FIG. 8 is a schematic of video synthesis according to an embodiment
of the invention;
FIG. 9 is a block diagram of a prior art reference picture
management;
FIG. 10 is a block diagram of multiview reference picture
management according to an embodiment of the invention;
FIG. 11 is a block diagram of multiview reference pictures in a
decoded picture buffer according to an embodiment of the
invention;
FIG. 12 is a graph comparing coding efficiencies of different
multiview reference picture orderings;
FIG. 13 is a block diagram of dependencies of view mode on the
multiview reference picture list manager according to an embodiment
of the invention;
FIG. 14 is a diagram of a prior art reference picture management
for single view coding systems that employ prediction from temporal
reference pictures;
FIG. 15 is a diagram of a reference picture management for
multiview coding and decoding systems that employ prediction from
multiview reference pictures according to an embodiment of the
invention;
FIG. 16 is a block diagram of view synthesis in a decoder using
depth information encoded and received as side information
according to an embodiment of the invention;
FIG. 17 is a block diagram of cost calculations for selecting a
prediction mode according to an embodiment of the invention;
FIG. 18 is a block diagram of view synthesis in a decoder using
depth information estimated by a decoder according to an embodiment
of the invention;
FIG. 19 is a block diagram of multiview videos using V-frames to
achieve spatial random access in the decoder according to an
embodiment of the invention;
FIG. 20 is a block diagram of view synthesizes using warping and
interpolation according to an embodiment of the invention; and
FIG. 21 is a block diagram of a depth search according to an
embodiment of the invention;
FIG. 22 is a block diagram of sub-pel reference matching according
to an embodiment of the invention:
FIG. 23 is a block diagram of conventional skip mode; and
FIG. 24 is a block diagram of synthetic skip mode according to an
embodiment of the invention.
DETAILED DESCRIPTION OF THE EMBODIMENTS OF THE INVENTION
One embodiment of our invention provides a joint
temporal/inter-view processing method for encoding and decoding
frames of multiview videos. Multiview videos are videos that are
acquired of a scene by multiple cameras having different poses. We
define a pose camera as both its 3D (x, y, z) position, and its 3D
(.theta., .rho., .phi.) orientation. Each pose corresponds to a
`view` of the scene.
The method uses temporal correlation between frames within the same
video acquired for a particular camera pose, as well as spatial
correlation between synchronized frames in different videos
acquired from multiple camera views. In addition, `synthetic`
frames can be correlated, as described below.
In one embodiment, the temporal correlation uses motion compensated
temporal filtering (MCTF), while the spatial correlation uses
disparity compensated inter-view filtering (DCVF).
In another embodiment of the invention, spatial correlation uses
prediction of one view from synthesized frames that are generated
from `neighboring` frames. Neighboring frames are temporally or
spatially adjacent frames, for example, frames before or after a
current frame in the temporal domain, or one or more frames
acquired at the same instant in time but from cameras having
different poses or views of the scene.
Each frame video includes macroblocks of pixels. Therefore, the
method of multiview video encoding and decoding according to one
embodiment of the invention is macroblock adaptive. The encoding
and decoding of a current macroblock in a current frame is
performed using several possible prediction modes, including
various forms of temporal, spatial, view synthesis, and intra
prediction. To determine the best prediction mode on a macroblock
basis, one embodiment of the invention provides a method for
selecting a prediction mode. The method can be used for any number
of camera arrangements.
As used herein, a reference picture is defined as any frame that is
used during the encoding and decoding to `predict` a current frame.
Typically, reference pictures are spatially or temporally adjacent
or `neighboring` to the current frame.
It is important to note that the same operations are applied in
both the encoder and decoder because the same set of reference
pictures are used at any given time instant to encode and decode
the current frame.
MCTF/DCVF Decomposition
FIG. 4 shows a MCTF/DCVF decomposition 400 according to one
embodiment of the invention. Frames of input videos 401-404 are
acquired of a scene 5 by cameras 1-4 having different posses. Note,
as shown in FIG. 8, some of the cameras 1a and 1b can be at the
same locations but with different orientations. It is assumed that
there is some amount of view overlap between any pair of cameras.
The poses of the cameras can change while acquiring the multiview
videos. Typically, the cameras are synchronized with each other.
Each input video provides a different `view` of the scene. The
input frames 401-404 are sent to a MCTF/DCVF decomposition 400. The
decomposition produces encoded low-band frames 411, encoded high
band frames 412, and associated side information 413. The high band
frames encode prediction errors using the low band frames as
reference pictures. The decomposition is according to selected
prediction modes 410. The prediction modes include spatial,
temporal, view synthesis, and intra prediction modes. The
prediction modes can be selected adaptively on a per macroblock
basis for each current frame. With intra prediction, the current
macroblock is predicted from other macroblocks in the same
frame.
FIG. 5 shows a preferred alternating `checkerboard pattern` of the
low band frames (L) 411 and the high band frames (H) 412 for a
neighborhood of frames 510. The frames have a spatial (view)
dimension 501 and a temporal dimension 502. Essentially, the
pattern alternates low band frames and high band frames in the
spatial dimension for a single instant in time, and additionally
alternates temporally the low band frames and the high band frames
for a single video.
There are several advantages of this checkerboard pattern. The
pattern distributes low band frames evenly in both the space and
time dimensions, which achieves scalability in space and time when
a decoder only reconstructs the low band frames. In addition, the
pattern aligns the high band frames with adjacent low band frames
in both the space and time dimensions. This maximizes the
correlation between reference pictures from which the predictions
of the errors in the current frame are made, as shown in FIG.
6.
According to a lifting-based wavelet transform, the high band
frames 412 are generated by predicting one set of samples from the
other set of samples. The prediction can be achieved using a number
of modes including various forms of temporal prediction, various
forms of spatial prediction, and a view synthesis prediction
according to the embodiments of invention described below.
The means by which the high band frames 412 are predicted and the
necessary information required to make the prediction are referred
to as the side information 413. If a temporal prediction is
performed, then the temporal mode is signaled as part of the side
information along with corresponding motion information. If a
spatial prediction is performed, then the spatial mode is signaled
as part of the side information along with corresponding disparity
information. If view synthesis prediction is performed, then the
view synthesis mode is signaled as part of the side information
along with corresponding disparity, motion and depth
information.
As shown in FIG. 6, the prediction of each current frame 600 uses
neighboring frames 510 in both the space and time dimensions. The
frames that are used for predicting the current frame are called
reference pictures. The reference pictures are maintained in the
reference list, which is part of the encoded bitstream. The
reference pictures are stored in the decoded picture buffer.
In one embodiment of the invention, the MCTF and DCVF are applied
adaptively to each current macroblock for each frame of the input
videos to yield decomposed low band frames, as well as the high
band frames and the associated side information. In this way, each
macroblock is processed adaptively according to a `best` prediction
mode. An optimal method for selecting the prediction mode is
described below.
In one embodiment of the invention, the MCTF is first applied to
the frames of each video independently. The resulting frames are
then further decomposed with the DCVF. In addition to the final
decomposed frames, the corresponding side information is also
generated. If performed on a macroblock-basis, then the prediction
mode selections for the MCTF and the DCVF are considered
separately. As an advantage, this prediction mode selection
inherently supports temporal scalability. In this way, lower
temporal rates of the videos are easily accessed in the compressed
bitstream.
In another embodiment, the DCVF is first applied to the frames of
the input videos. The resulting frames are then temporally
decomposed with the MCTF. In addition to the first decomposed
frames, the side information is also generated. If performed on a
macroblock-basis, then the prediction mode selections for the MCTF
and DCVF are considered separately. As an advantage, this selection
inherently supports spatial scalability. In this way, a reduced
number of the views are easily accessed in the compressed
bitstream.
The decomposition described above can be applied recursively on the
resulting set of low band frames from a previous decomposition
state. As an advantage, our MCTF/DCVF decomposition 400 effectively
removes both temporal and spatial (inter-view) correlations, and
can achieve a very high compression efficiency. The compression
efficiency of our multiview video encoder outperforms conventional
simulcast encoding, which encodes each video for each view
independently.
Encoding of MCTF/DCVF Decomposition
As shown in FIG. 7, the outputs 411 an 412 of decomposition 400 are
fed to a single encoder 710, and the output 413 is fed to a side
information encoder 720. The signal encoder 710 performs a
transform, quantization and entropy coding to remove any remaining
correlations in the decomposed low band and high band frames
411-412. Such operations are well known in the art. Netravali and
Haskell, Digital Pictures: Representation, Compression and
Standards, Second Edition, Plenum Press, 1995.
The side information encoder 720 encodes the side information 413
generated by the decomposition 400. In addition to the prediction
mode and the reference picture list, the side information 413
includes motion information corresponding to the temporal
predictions, disparity information corresponding to the spatial
predictions and view synthesis and depth information corresponding
to the view synthesis predictions.
Encoding the side information can be achieved by known and
established techniques, such as the techniques used in the MPEG-4
Visual standard, ISO/IEC 14496-2, "Information technology--Coding
of audio-visual objects--Part 2: Visual," 2.sup.nd Edition, 2001,
or the more recent H.264/AVC standard, and ITU-T Recommendation
H.264, "Advanced video coding for generic audiovisual services,"
2004.
For instance, motion vectors of the macroblocks are typically
encoded using predictive methods that determine a prediction vector
from vectors in macroblocks in reference pictures. The difference
between the prediction vector and the current vector is then
subject to an entropy coding process, which typically uses the
statistics of the prediction error. A similar procedure can be used
to encode disparity vectors.
Furthermore, depth information for each macroblock can be encoded
using predictive coding methods in which a prediction from
macroblocks in reference pictures is obtained, or by simply using a
fixed length code to express the depth value directly. If pixel
level accuracy for the depth is extracted and compressed, then
texture coding techniques that apply transform, quantization and
entropy coding techniques can be applied.
The encoded signals 711-713 from the signal encoder 710 and side
information encoder 720 can be multiplexed 730 to produce an
encoded output bitstream 731.
Decoding of MCTF/DCVF Decomposition
The bitstream 731 can be decoded 740 to produce output multiview
videos 741 corresponding to the input multiview videos 401-404.
Optionally, synthetic video can also be generated. Generally, the
decoder performs the inverse operations of the encoder to
reconstruct the multiview videos. If all low band and high band
frames are decoded, then the full set of frames in both the space
(view) dimension and time dimension at the encoded quality are
reconstructed and available.
Depending on the number of recursive levels of decomposition that
were applied in the encoder and which type of decompositions were
applied, a reduced number of videos and/or a reduced temporal rate
can be decoded as shown in FIG. 7.
View Synthesis
As shown in FIG. 8, view synthesis is a process by which frames 801
of a synthesized video are generated from frames 803 of one or more
actual multiview videos. In other words, view synthesis provides a
means to synthesize the frames 801 corresponding to a selected
novel view 802 of the scene 5. This novel view 802 may correspond
to a `virtual` camera 800 not present at the time the input
multiview videos 401-404 were acquired or the view can correspond
to a camera view that is acquired, whereby the synthesized view
will be used for prediction and encoding/decoding of this view as
described below.
If one video is used, then the synthesis is based on extrapolation
or warping, and if multiple videos are used, then the synthesis is
based on interpolation.
Given the pixel values of frames 803 of one or more multiview
videos and the depth values of points in the scene, the pixels in
the frames 801 for the synthetic view 802 can be synthesized from
the corresponding pixel values in the frames 803.
View synthesis is commonly used in computer graphics for rendering
still images for multiple views, see Buehler et al., "Unstructured
Lumigraph Rendering," Proc. ACM SIGGRAPH, 2001. That method
requires extrinsic and intrinsic parameters for the cameras,
incorporated herein by reference
View synthesis for compressing multiview videos is novel. In one
embodiment of our invention, we generate synthesized frames to be
used for predicting the current frame. In one embodiment of the
invention, synthesized frames are generated for designated high
band frames. In another embodiment of the invention, synthesized
frames are generated for specific views. The synthesized frames
serve as reference pictures from which a current synthesized frame
can be predicted.
One difficulty with this approach is that the depth values of the
scene 5 are unknown. Therefore, we estimate the depth values using
known techniques, e.g., based on correspondences of features in the
multiview videos.
Alternatively, for each synthesized video, we generate multiple
synthesized frames, each corresponding to a candidate depth value.
For each macroblock in the current frame, the best matching
macroblock in the set of synthesized frames is determined. The
synthesized frame from which this best match is found indicates the
depth value of the macroblock in the current frame. This process is
repeated for each macroblock in the current frame.
A difference between the current macroblock and the synthesized
block is encoded and compressed by the signal encoder 710. The side
information for this multiview is encoded by the side information
encoder 720. This side information includes a signal indicating the
view synthesis prediction mode, the depth value of the macroblock,
and an optional displacement vector that compensates for any
misalignments between the macroblock in the current frame and the
best matching macroblock in the synthesized frame to be
compensated.
Prediction Mode Selection
In the macroblock-adaptive MCTF/DCVF decomposition, the prediction
mode m for each macroblock can be selected by minimizing a cost
function adaptively on a per macroblock basis:
.times..times..times..times..function. ##EQU00001##
where J(m)=D(m)+.lamda.R(m), and D is distortion, .lamda. is a
weighting parameter, R is rate, m indicates the set of candidate
prediction modes, and m* indicates the optimal prediction mode that
has been selected based on a minimum cost criteria.
The candidate modes m include various modes of temporal, spatial,
view synthesis, and intra prediction. The cost function J(m)
depends on the rate and distortion resulting from encoding the
macroblock using a specific prediction mode m.
The distortion D measures a difference between a reconstructed
macroblock and a source macroblock. The reconstructed macroblock is
obtained by encoding and decoding the macroblock using the given
prediction mode m. A common distortion measure is a sum of squared
difference. The rate R corresponds to the number of bits needed to
encode the macroblock, including the prediction error and the side
information. The weighting parameter .lamda. controls the
rate-distortion tradeoff of the macroblock coding, and can be
derived from a size of a quantization step.
Detailed aspects of the encoding and decoding processes are
described in further detail below. In particular, the various data
structures that are used by the encoding and decoding processes are
described. It should be understood that the data structures, as
described herein, that are used in the encoder are identical to
corresponding data structures used in the decoder. It should also
be understood that the processing steps of the decoder essentially
follow the same processing steps as the encoder, but in an inverse
order.
Reference Picture Management
FIG. 9 shows a reference picture management for prior art
single-view encoding and decoding systems. Temporal reference
pictures 901 are managed by a single-view reference picture list
(RPL) manager 910, which determines insertion 920 and removal 930
of temporal reference pictures 901 to a decoded picture buffer
(DBP) 940. A reference picture list 950 is also maintained to
indicate the frames that are stored in the DPB 940. The RPL is used
for reference picture management operations such as insert 920 and
remove 930, as well as temporal prediction 960 in both the encoded
and the decoder.
In single-view encoders, the temporal reference pictures 901 are
generated as a result of applying a set of typical encoding
operations including prediction, transform and quantization, then
applying the inverse of those operations including inverse
quantization, inverse transform and motion compensation.
Furthermore, temporal reference pictures 901 are only inserted into
the DPB 940 and added to the RPL 950 when the temporal pictures are
required for the prediction of a current frame in the encoder.
In single-view decoders, the same temporal reference picture 901
are generated by applying a set of typical decoding operations on
the bitstream including inverse quantization, inverse transform and
motion compensation. As in the encoder, the temporal reference
pictures 901 are only inserted 920 into the DPB 940 and added to
the RPL 950 if they are required for prediction of a current frame
in the decoder.
FIG. 10 shows a reference picture management for multiview encoding
and decoding. In addition to temporal reference pictures 1003, the
multiview systems also include spatial reference pictures 1001 and
synthesized reference pictures 1002. These reference pictures are
collectively referred to as multiview reference pictures 1005. The
multiview reference pictures 1005 are managed by a multiview RPL
manager 1010, which determines insertion 1020 and removal 1030 of
the multiview reference pictures 1005 to the multiview DPB 1040.
For each video, a multiview reference picture list (RPL) 1050 is
also maintained to indicate the frames that are stored in the DPB.
That is, the RPL is an index for the DPB. The multiview RPLs are
used for reference picture management operations such as insert
1020 and remove 1030, as well as prediction 1060 of the current
frame.
It is noted that prediction 1060 for the multiview system is
different than prediction 960 for the single-view system because
prediction from different types of multiview reference pictures
1005 is enabled. Further details on the multiview reference picture
management 1010 are described below.
Multiview Reference Picture List Manager
Before encoding a current frame in the encoder or before decoding
the current frame in the decoder, a set of multiview reference
pictures 1005 can be indicated in the multiview RPL 1050. As
defined conventionally and herein, a set can have zero (null set),
one or multiple elements. Identical copies of the RPLs are
maintained by both the encoder and decoder for each current
frame.
All frames inserted in the multiview RPLs 1050 are initialized and
marked as usable for prediction using an appropriate syntax.
According to the H.264/AVC standard and reference software, the
`used_for_reference` flag is set to `1`. In general, reference
pictures are initialized so that a frame can be used for prediction
in a video encoding system. To maintain compatability with
conventional single-view video compression standards, such as
H.264/AVC, each reference picture is assigned a picture order count
(POC). Typically, for single-view encoding and decoding systems,
the POC corresponds to the temporal ordering of a picture, e.g.,
the frame number. For multiview encoding and decoding systems,
temporal order alone is not sufficient to assign a POC for each
reference picture. Therefore, we determine a unique POC for every
multiview reference picture according to a convention. One
convention is to assign a POC for temporal reference pictures based
on temporal order, and then to reserve a sequence of very high POC
numbers, e.g., 10,000-10,100, for the spatial and synthesized
reference pictures. Other POC assignment conventions, or simply
"ordering" conventions, are described in further detail below.
All frames used as multiview reference pictures are maintained in
the RPL and stored in the DPB in such a way that the frames are
treated as conventional reference pictures by the encoder 700 or
the decoder 740. This way, the encoding and decoding processes can
be conventional. Further details on storing multiview reference
pictures are described below. For each current frame to be
predicted, the RPL and DPB are updated accordingly.
Defining and Signaling Multiview Conventions
The process of maintaining the RPL is coordinated between the
encoder 700 and the decoder 740. In particular, the encoder and
decoder maintain identical copies of multiview reference picture
list when predicting a particular current frame.
A number of conventions for maintaining the multi-frame reference
picture list are possible. Therefore, the particular convention
that is used is inserted in the bitstream 731, or provided as
sequence level side information, e.g., configuration information
that is communicated to the decoder. Furthermore, the convention
allows different prediction structures, e.g., 1-D arrays, 2-D
arrays, arcs, crosses, and sequences synthesized using view
interpolation or warping techniques.
For example, a synthesized frame is generated by warping a
corresponding frame of one of the multiview videos acquired by the
cameras. Alternatively, a conventional model of the scene can be
used during the synthesis. In other embodiments of our invention,
we define several multiview reference picture maintenance
conventions that are dependent on view type, insertion order, and
camera properties.
The view type indicates whether the reference picture is a frame
from a video other than the video of the current frame, or whether
the reference picture is synthesized from other frames, or whether
the reference picture depends on other reference pictures. For
example, synthesized reference pictures can be maintained
differently than reference pictures from the same video as the
current frame, or reference pictures from spatially adjacent
videos.
The insertion order indicates how reference pictures are ordered in
the RPL. For instance, a reference picture in the same video as the
current frame can be given a lower order value than a reference
picture in a video taken from an adjacent view. In this case, the
reference picture is placed earlier in the multiview RPL.
Camera properties indicate properties of the camera that is used to
acquire the reference picture, or the virtual camera that is used
to generate a synthetic reference picture. These properties include
translation and rotation relative to a fixed coordinate system,
i.e., the camera `pose`, intrinsic parameters describing how a 3-D
point is projected into a 2-D image, lens distortions, color
calibration information, illumination levels, etc. For instance,
based on the camera properties, the proximity of certain cameras to
adjacent cameras can be determined automatically, and only videos
acquired by adjacent cameras are considered as part of a particular
RPL.
As shown in FIG. 11, one embodiment of our invention uses a
convention that reserves a portion 1101 of each reference picture
list for temporal reference pictures 1003, reserves another portion
1102 for synthesized reference pictures 1002 and a third portion
1103 for spatial reference pictures 1001. This is an example of a
convention that is dependent only on the view type. The number of
frames contained in each portion can vary based on a prediction
dependency of the current frame being encoded or decoded.
The particular maintenance convention can be specified by standard,
explicit or implicit rules, or in the encoded bitstream as side
information.
Storing Pictures in the DPB
The multiview RPL manager 1010 maintains the RPL so that the order
in which the multiview reference pictures are stored in the DPB
corresponds to their `usefulness` to improve the efficiency of the
encoding and decoding. Specifically, reference pictures in the
beginning of the RPL can be predicatively encoded with fewer bits
than reference pictures at the end of the RPL.
As shown in FIG. 12, optimizing the order in which multiview
references pictures are maintained in the RPL can have a
significant impact on coding efficiency. For example, following the
POC assignment described above for initialization, multiview
reference pictures can be assigned a very large POC value because
they do not occur in the normal temporal ordering of a video
sequence. Therefore, the default ordering process of most video
codecs can place such multiview reference pictures earlier in the
reference picture lists.
Because temporal reference pictures from the same sequence
generally exhibit stronger correlations than spatial reference
pictures from other sequences, the default ordering is undesirable.
Therefore, the multiview reference pictures are either explicitly
reordered by the encoder, whereby the encoder then signals this
reordering to the decoder, or the encoder and decoder implicitly
reorder multiview reference pictures according to a predetermined
convention.
As shown in FIG. 13, the order of the reference pictures is
facilitated by a view mode 1300 to each reference picture. It is
noted that the view mode 1300 also affects the multiview prediction
process 1060. In one embodiment of our invention, we use three
different types of view modes, I-view, P-view and B-view, which are
described in further detail below.
Before describing the detailed operation of multiview reference
picture management, prior art reference picture management for
single video encoding and decoding systems is shown in FIG. 14.
Only temporal reference pictures 901 are used for the temporal
prediction 960. The temporal prediction dependency between temporal
reference pictures of the video in acquisition or display order
1401 is shown. The reference pictures are reordered 1410 into an
encoding order 1402, in which each reference picture is encoded or
decoded at a time instants t.sub.0-t.sub.6. Block 1420 shows the
ordering of the reference pictures for each instant in time. At
time t.sub.0, when an intra-frame I.sub.0 is encoded or decoded,
there are no temporal reference pictures are used for temporal
prediction, hence the DBP/RPL is empty. At time t.sub.1, when the
uni-directional inter-frame P.sub.1 is encoded or decoded, frame
I.sub.0 is available as a temporal reference picture. At times
t.sub.2 and t.sub.3, both frames I.sub.0 and P.sub.1 are available
as reference frames for bi-directional temporal prediction of
inter-frames B.sub.1 and B.sub.2. The temporal reference pictures
and DBP/RPL are managed in a similar way for future pictures.
To describe the multiview case according to an embodiment of the
invention, we consider the three different types of views described
above and shown in FIG. 15: I-view, P-view, and B-view. The
multiview prediction dependency between reference pictures of the
videos in display order 1501 is shown. As shown in FIG. 15, the
reference pictures of the videos are reordered 1510 into a coding
order 1502 for each view mode, in which each reference picture is
encoded or decoded at a given time instant denoted t.sub.0-t.sub.2.
The order of the multiview reference pictures is shown in block
1520 for each time instant.
The I-view is the simplest mode that enables more complex modes.
I-view uses conventional encoding and prediction modes, without any
spatial or synthesized prediction. For example, I-views can be
encoded using conventional H.264/AVC techniques without any
multiview extensions. When spatial reference pictures from an
I-view sequence are placed into the reference lists of other views,
these spatial reference pictures are usually placed after temporal
reference pictures.
As shown in FIG. 15, for the I-view, when frame I.sub.0 is encoded
or decoded at t.sub.0, there are no multiview reference pictures
used for prediction. Hence the DBP/RPL is empty. At time t.sub.1,
when frame P.sub.0 is encoded or decoded, I.sub.0 is available as a
temporal reference picture. At time t.sub.2, when the frame B.sub.0
is encoded or decoded, both frames I.sub.0 and P.sub.0 are
available as temporal reference pictures.
P-view is more complex than I-view in that P-view allows prediction
from another view to exploit the spatial correlation between views.
Specifically, sequences encoded using the P-view mode use multiview
reference pictures from other I-view or P-view. Synthesized
reference pictures can also be used in the P-view. When multiview
reference pictures from an I-view are placed into the reference
lists of other views, P-views are placed after both temporal
reference pictures and after multiview references pictures derived
from I-views.
As shown in FIG. 15, for the P-view, when frame I.sub.2 is encoded
or decoded at t.sub.0, a synthesized reference picture S.sub.20 and
the spatial reference picture I.sub.0 are available for prediction.
Further details on the generation of synthesized pictures are
described below. At time t.sub.1, when P.sub.2 is encoded or
decoded, I.sub.2 is available as a temporal reference picture,
along with a synthesized reference picture S.sub.21 and a spatial
reference picture P.sub.0 from the I-view. At time t.sub.2, there
exist two temporal reference pictures I.sub.2 and P.sub.2, as well
as a synthesized reference picture S.sub.22 and a spatial reference
picture B.sub.0, from which predictions can be made.
B-views are similar to P-views in that the B-views use multiview
reference pictures. One key difference between P-views and B-views
is that P-views use reference pictures from its own view as well as
one other view, while B-views may reference pictures in multiple
views. When synthesized reference pictures are used, the B-views
are placed before spatial reference pictures because synthesized
views generally have a stronger correlation than spatial
references.
As shown in FIG. 15, for the B-view, when I.sub.1 is encoded or
decoded at t.sub.0, a synthesized reference picture S.sub.10 and
the spatial reference pictures I.sub.0 and I.sub.2 are available
for prediction. At time t.sub.1, when P.sub.1 is encoded or
decoded, I.sub.1 is available as a temporal reference picture,
along with a synthesized reference picture S.sub.11 and spatial
reference pictures P.sub.0 and P.sub.2 from the I-view and P-view,
respectively. At time t.sub.2, there exist two temporal reference
pictures I.sub.1 and P.sub.1, as well as a synthesized reference
picture S.sub.12 and spatial reference pictures B.sub.0 and
B.sub.2, from which predictions can be made.
It must be emphasized that the example shown in FIG. 15 is only for
one embodiment of the invention. Many different types of prediction
dependencies are supported. For instance, the spatial reference
pictures are not limited to pictures in different views at the same
time instant. Spatial reference pictures can also include reference
pictures for different views at different time instants. Also, the
number of bi-directionally predicted pictures between
intra-pictures and uni-directionally predicted inter-pictures can
vary. Similarly, the configuration of I-views, P-views, and B-views
can also vary. Furthermore, there can be several synthesized
reference pictures available, each generated using a different set
of pictures or different depth map or process.
Compatibility
One important benefit of the multiview picture management according
to the embodiments of the invention is that it is compatible with
existing single-view video coding systems and designs. Not only
does this provide minimal changes to the existing single-view video
coding standards, but it also enables software and hardware from
existing single view video coding systems to be used for multiview
video coding as described herein.
The reason for this is that most conventional video encoding
systems communicate encoding parameters to a decoder in a
compressed bitstream. Therefore, the syntax for communicating such
parameters is specified by the existing video coding standards,
such as the H.264/AVC standard. For example, the video coding
standard specifies a prediction mode for a given macroblock in a
current frame from other temporally related reference pictures. The
standard also specifies methods used to encode and decode a
resulting prediction error. Other parameters specify a type of size
of a transform, a quantization method, and an entropy coding
method.
Therefore, our multiview reference pictures can be implemented with
only limited number of modifications to standard encoding and
decoding components such as the reference picture lists, decoded
picture buffer, and prediction structure of existing systems. It is
noted that the macroblock structure, transforms, quantization and
entropy encoding remain unchanged.
View Synthesis
As described above for FIG. 8, view synthesis is a process by which
frames 801 corresponding to a synthetic view 802 of a virtual
camera 800 are generated from frames 803 acquired of existing
videos. In other words, view synthesis provides a means to
synthesize the frames corresponding to a selected novel view of the
scene by a virtual camera not present at the time the input videos
were acquired. Given the pixel values of frames of one or more
actual video and the depth values of points in the scene, the
pixels in the frames of the synthesized video view can be generated
by extrapolation and/or interpolation.
Prediction from Synthesized Views
FIG. 16 shows a process for generating a reconstructed macroblock
using the view-synthesis mode, when depth 1901 information is
included in the encoded multiview bitstream 731. The depth for a
given macroblock is decoded by a side information decoder 1910. The
depth 1901 and the spatial reference pictures 1902 are used to
perform view synthesis 1920, where a synthesized macroblock 1904 is
generated. A reconstructed macroblock 1903 is then formed by adding
1930 the synthesized macroblock 1904 and a decoded residual
macroblock 1905.
Details on Multiview Mode Selection at Encoder
FIG. 17 shows a process for selecting the prediction mode while
encoding or decoding a current frame. Motion estimation 2010 for a
current macroblock 2011 is performed using temporal reference
pictures 2020. The resultant motion vectors 2021 are used to
determine 2030 a first coding cost, cost.sub.1 2031, using temporal
prediction. The prediction mode associated with this process is
m.sub.1.
Disparity estimation 2040 for the current macroblock is performed
using spatial reference pictures 2041. The resultant disparity
vectors 2042 are used to determine 2050 a second coding cost,
cost.sub.2 2051, using spatial prediction. The prediction mode
associated with this process is denoted m.sub.2.
Depth estimation 2060 for the current macroblock is performed based
on the spatial reference pictures 2041. View synthesis is performed
based on the estimated depth. The depth information 2061 and the
synthesized view 2062 are used to determine 2070 a third coding
cost, cost.sub.3 2071, using view-synthesis prediction. The
prediction mode associated this process is m.sub.3.
Adjacent pixels 2082 of the current macroblock are used to
determine 2080 a fourth coding cost, cost.sub.4 2081, using intra
prediction. The prediction mode associated with process is
m.sub.4.
The minimum cost among cost.sub.1, cost.sub.2, cost.sub.3 and
cost.sub.4 is determined 2090, and one of the modes m.sub.1,
m.sub.2, m.sub.3 and m.sub.4 that has the minimum cost is selected
as the best prediction mode 2091 for the current macroblock
2011.
View Synthesis Using Depth Estimation
Using the view synthesis mode 2091, the depth information and
displacement vectors for synthesized views can be estimated from
decoded frames of one or more multiview videos. The depth
information can be per-pixel depth estimated from stereo cameras,
or it can be per-macroblock depth estimated from macroblock
matching, depending on the process applied.
An advantage of this approach is a reduced bandwidth because depth
values and displacement vectors are not needed in the bitstream, as
long as the encoded has access to the same depth and displacement
information as the decoder. The encoder can achieve this as long as
the decoder uses exactly the same depth and displacement estimation
process as the encoder. Therefore, in this embodiment of the
invention, a difference between the current macroblock and the
synthesized macroblock is encoded by the encoder.
The side information for this mode is encoded by the side
information encoder 720. The side information includes a signal
indicating the view synthesis mode and the reference view(s). The
side information can also include depth and displacement correction
information, which is the difference between the depth and
displacement used by the encoder for view synthesis and the values
estimated by the decoder.
FIG. 18 shows the decoding process for a macroblock using the
view-synthesis mode when the depth information is estimated or
inferred in the decoder and is not conveyed in the encoded
multiview bitstream. The depth 2101 is estimated 2110 from the
spatial reference pictures 2102. The estimated depth and the
spatial reference pictures are then used to perform view synthesis
2120, where a synthesized macroblock 2121 is generated. A
reconstructed macroblock 2103 is formed by the addition 2130 of the
synthesized macroblock and the decoded residual macroblock
2104.
Spatial Random Access
In order to provide random access to frames in a conventional
video, intra-frames, also known as I-frames, are usually spaced
throughout the video. This enables the decoder to access any frame
in the decoded sequence, although at a decreased compression
efficiency.
For our multiview encoding and decoding system, we provide a new
type of frame, which we call a `V-frame` to enable random access
and increase compression efficiency. A V-frame is similar to an
I-frame in the sense that the V-frame is encoded without any
temporal prediction. However, the V-frame also allows prediction
from other cameras or prediction from synthesized videos.
Specifically, V-frames are frames in the compressed bitstream that
are predicted from spatial reference pictures or synthesized
reference pictures. By periodically inserting V-frames, instead of
I-frames, in the bitstream, we provide temporal random access as is
possible with I-frames, but with a better encoding efficiency.
Therefore, V-frames do not use temporal reference frames. FIG. 19
shows the use of I-frames for the initial view and the use of
V-frames for subsequent views at the same time instant 1900. It is
noted that for the checkerboard configuration shown in FIG. 5,
V-frames would not occur at the same time instant for all views.
Any of the low-band frames could be assigned a V-frame. In this
case, the V-frames would be predicted from low-band frames of
neighboring views.
In H.264/AVC video coding standard, IDR frames, which are similar
to MPEG-2 I-frames with closed GOP, imply that all reference
pictures are removed from the decoder picture buffer. In this way,
the frame before an IDR cannot be used to predict frames after the
IDR frame.
In the multiview decoder as described herein, V-frames similarly
imply that all temporal reference pictures can be removed from the
decoder picture buffer. However, spatial reference pictures can
remain in the decoder picture buffer. In this way, a frame in a
given view before the V-frame cannot be used to perform temporal
prediction for a frame in the same view after the V-frame.
To gain access to a particular frame in one of the multiview
videos, the V-frame for that view must first be decoded. As
described above, this can be achieved through prediction from
spatial reference pictures or synthesized reference pictures,
without the use of temporal reference pictures.
After the V-frame of the select view is decoded, subsequent frames
in that view are decoded. Because these subsequent frames are
likely to have a prediction dependency on reference pictures from
neighboring views, the reference pictures in these neighboring
views are also be decoded.
Multiview Encoding and Decoding
The above sections describe view synthesis for improved prediction
in multiview coding and depth estimation. We now describe
implementation for variable block-size depth and motion search,
rate-distortion (RD) decision, sub-pel reference depth search, and
context-adaptive binary arithmetic coding (CABAC) of depth
information. The coding can include encoding in an encoder and
decoding in a decoder. CABAC is specified by the H.624 standard
PART 10, incorporated herein by reference.
View Synthesis Prediction
To capture the correlations that exist both across cameras and
across time, we implemented two methods of block prediction:
1) disparity compensated view prediction (DCVP); and
2) view synthesis prediction (VSP).
DCVP
The first method, DCVP, corresponds to using a frame from a
different camera (view) at the same time to predict a current
frame, instead of using a frame from a different time of the same
(view) camera. DCVP provides gains when a temporal correlation is
lower than the spatial correlation, e.g., due to occlusions,
objects entering or leaving the scene, or fast motion.
VSP
The second method, VSP, synthesizing frames for a virtual camera to
predict a sequence of frames. VSP is complementary to DCVP due to
the existence of non-translational motion between camera views an
provide gains when the camera parameters are sufficiently accurate
to provide high quality virtual views, which is often the case in
real applications.
As shown in FIG. 20, we exploit these features of multiview video
by synthesizing a virtual view from already encoded views and then
performing predictive coding using the synthesized views. FIG. 20
shows time on the horizontal axis, and views on the vertical axis,
with view synthesis and warping 2001, and view synthesis and
interpolation 2002.
Specifically, for each camera c, we first synthesize a virtual
frame I'[c, t, x, y] based on the unstructured lumigraph rendering
technique of Buchler et al., see above, and then predicatively
encode the current sequence using the synthesized view.
To synthesize a frame I'[c, t, x, y], we require a depth map D[c,
t, x, y] that indicate how far the object corresponding to pixel
(x, y) is from camera c at time t, as well as an intrinsic matrix
A(c), rotation matrix R(c), and a translation vector T(c)
describing the location of camera c relative to some world
coordinate system.
Using these quantities, we can apply the well-known pinhole camera
model to project the pixel location (x, y) into world coordinates
[u, v, w] via [u,v,w]=R(c)A.sup.-1(c)[x,y,1]D[c,t,x,y]+T(c).
(1)
Next, the world coordinates are mapped to a target coordinates [x',
y', z'] of the frame in camera c', which we wish to predict from
via [x',y',z']=A(c')R.sup.-1(c'){[u,v,w]-T(c')}. (2)
Finally, to obtain a pixel location, the target coordinates are
converted to homogenous from [x'/z', y'/z', l], and an intensity
for pixel location (x, y) in the synthesized frame is I'[c, t, x,
y]=I[c',t, x'/z', y'/z'].
Variable Block-Size Depth/Motion Estimation
Above, we described a picture buffer-management method that enables
for the use of DCVP without changing the syntax. The disparity
vectors between camera views were found by using the motion
estimation step and could be used just as an extended reference
type. In order to use VSP as another type of reference, we extend
the typical motion-estimation process as follows.
Given a candidate macroblock type mb_type and N possible reference
frames, possibly including synthesized multiview reference frames,
i.e., VSP, we find, for each sub-macroblock, a reference frame
along with either a motion vector rar m or a
depth/correction-vector pair (d,{right arrow over (m.sub.c)}) that
minimizes a following Lagrangian cost J using a Lagrange multiplier
.lamda..sub.motion or .lamda..sub.depth, respectively,
J=min(J.sub.motion,J.sub.depth), with
.function..fwdarw..lamda..di-elect
cons..times..times..times..times..times..function..fwdarw..lamda..functio-
n..times..times..times..times..times..times..times..function..fwdarw..lamd-
a..di-elect
cons..times..times..times..times..times..function..fwdarw..lamda..functio-
n. ##EQU00002## where the sum is taken over all the pixels in the
sub-macroblock being considered (sub-MB), and
X.sub.p.sub.--.sub.synth or X.sub.p.sub.--.sub.motion refer to the
intensity of a pixel in the reference sub-macroblock.
Note that here `motion` not only refers to temporal motion but also
inter-view motion resulting from the disparity between the
views.
Depth Search
We use a block-based depth search process to find the optimal depth
for each variable-size sub-macroblock. Specifically, we define
minimum, maximum, and incremental depth values D.sub.min,
D.sub.max, D.sub.step. Then, for each variable-size sub-macroblock
in the frame we wish to predict, we select the depth to minimize
the error for the synthesized block
.function..di-elect
cons..times..times..times..times..function..times.'''' ##EQU00003##
where .parallel.I[c, t, x, y]-I[c', t, x', y'].parallel. denotes an
average error between the sub-macroblock centered at (x, y) in
camera c at time t, and the corresponding block from which we
predict.
As an additional refinement to improve the performance of the basic
VSP process, we find that due to slight inaccuracies in the camera
parameters, non-idealities that are not captured by the pin hole
camera model, adding a synthesis correction vector significantly
improved the performance of VSP.
Specifically as shown in FIG. 21 for each macroblock 2100, we map
the target frame 2101 to the reference frame 2102, and then to the
synthesized frame 2103. However, instead of computing the
coordinates to interpolate from using Equation (1), we compute [u,
v, w] by adding a synthesis correction vector (C.sub.x,C.sub.y)
2110 to each set of original pixel coordinates to obtain
[u,v,w]=R(v)A.sup.-1(v)[x-C.sub.x,y-C.sub.y,l]D[v,t,x-C.sub.x,y-C.sub.y]+-
T(v). (4)
We discovered that with a correction-vector search range of as
small as +/-2 often significantly improves the quality of the
resulting synthesized reference frame.
Sub-Pixel Reference Matching
Because the disparity of two corresponding pixels in different
cameras is, in general, not given by an exact multiple of integers,
the target coordinates [x', y', z'] of the frame in camera c' which
we wish to predict from given by Equation (2) does not always fall
on an integer-grid point. Therefore, we use interpolation to
generate pixel-values for sub-pel positions in the reference frame.
This enables us to select a nearest sub-pel reference point,
instead of integer-pel, thereby more accurately approximating the
true disparity between the pixels.
FIG. 22 shows this process, where "oxx . . . ox" indicate pixels.
The same interpolation filters adopted for sub-pel motion
estimation in the H.264 standard are used in our
implementation.
Sub-Pixel Accuracy Correction Vector
We can further improve the quality of synthesis by allowing the use
of sub-pel accuracy correction vector. This is especially true when
combined with the sub-pel reference matching described above. Note
that there is a subtle difference between the sub-pel motion-vector
search and the current sub-pel correction-vector search.
In motion-vector cases, we usually search through the sub-pel
positions in the reference picture and selects a sub-pel
motion-vector pointing to the sub-pel position minimizing the RD
cost. However, in correction-vector cases, after finding the
optimal depth-value, we search through the sub-pel positions in the
current picture and select the correction-vector minimizing the RD
cost.
Shifting by a sub-pel correction-vector in the current picture does
not always lead to the same amount of shift in the reference
picture. In other words, the corresponding match in the reference
picture is always found by rounding to a nearest sub-pel position
after geometric transforms of Equations (1) and (2).
Although coding a sub-pel accuracy correction-vector is relatively
complex, we observe that the coding significantly improves the
synthesis quality and often leads to improved RD performance.
YUV-Depth Search
In depth estimation, regularization can achieve smoother depth
maps. Regularization improves visual quality of synthesized
prediction but slightly degrades its prediction quality as measured
by sum of absolute differences (SAD).
The conventional depth search process uses only the Y luminance
component of input images to estimate the depth in the depth map.
Although this minimizes the prediction error for Y component, if
often results in visual artifacts, e.g., in the form of color
mismatch, in the synthesized prediction. This means that we are
likely to have a degraded objective, i.e., PSNRs for U, V, as well
as subjective quality, in the form of color mismatch, in the final
reconstruction.
In order to cope with this problem, we extended the depth search
process to use the Y luminance component and the U and V
chrominance components. Using only the Y component can lead to
visual artifacts because a block can find a good match in a
reference frame by minimizing the prediction error, but these two
matching areas could be in two totally different colors. Therefore,
the quality of U and V prediction and reconstruction can improve by
incorporating the U and V components in the depth-search
process.
RD Mode Decision
A mode decision can be made by selecting the mb_type that minimizes
a Lagrangian cost function J.sub.mode defined as
.function..lamda..di-elect
cons..times..rho..lamda..function..times..times. ##EQU00004## where
X.sub.p refers to the value of a pixel in a reference MB, i.e.,
either a MB of a synthesized multiview, pure multiview or temporal
reference, and R.sub.side-info includes the bit rate for encoding
the reference index and the depth value/correction-value, or the
motion-vector depending on the type of the reference frame.
CABAC Encoding of Side-Information
Note that we have to encode a depth-value and a correction vector
for each synthesized MB when the MB is selected to be the best
reference via RD mode-decision. Both the depth value and the
correction vector are quantized using concatenated
unary/3.sup.rd-order Exp-Golomb (UEG3) binarization with signed
ValFlag=1, and cut off parameter uCoff=9 in exactly the same manner
as with the motion-vector.
Then, different context-models are assigned to the bins of the
resulting binary representation. The assignment of ctxIdxInc for
the depth and the correction-vector components is essentially the
same as that for motion-vectors as specified in Table 9-30 of ITU-T
Recommendation H.264 & ISO/IEC 14496-10 (MPEG-4) AVC, "Advanced
Video Coding for Generic Audiovisual Services," version 3: 2005,
incorporated herein by reference, except that we do not apply
sub-clause 9.3.3.1.1.7 to the first bin.
In our embodiment, we predicatively encode the depth value and the
correction vector using the same prediction scheme for
motion-vectors. Because a MB or a sub-MB of size down to 8.times.8
can have its own reference picture either from a temporal,
multiview, or synthesized multiview frame, the types of
side-information can vary from MB to MB. This implies the number of
neighboring MBs with identical reference picture can be reduced,
potentially leading to less efficient side-information
(motion-vector or depth/correction-vector) prediction.
When a sub-MB is selected to use a synthesized reference, but has
no surrounding MBs with the same reference, its
depth/correction-vector is independently coded without any
prediction. In fact, we found that it is often good enough to
binarize the correction-vector components with a fixed-length
representation followed by CABAC encoding of the resulting bins.
This is because MBs selected to use synthesized references tend to
be isolated, i.e., the MBs do not have neighboring MBs with the
same reference picture, and correction-vectors are usually less
correlated with their neighbors than in the case of motion-vector
cases.
Syntax and Semantics
As described above, we incorporate a synthesized reference picture,
in addition to the temporal and pure multiview references. Above,
we described a multiview reference picture list management method
that is compatible with the existing reference picture list
management in the H.264/AVC standard, referenced above.
The synthesized references in this embodiment are regarded as a
special case of multiview reference, and thus, are handled in
exactly the same way.
We define a new high level syntax element called a
view_parameter_set to describe the multiview identification and
prediction structure. By slightly modifying the parameter, we can
identify whether the current reference picture is of synthetic type
or not. Hence, we can either decode the depth/correction-vector or
the motion-vector for the given (sub-)MB depending on the type of
the reference. Therefore, we can integrate the use of this new type
of prediction by extending the macroblock-level syntax as specified
in Appendix A.
Skip Mode
In conventional skip mode, motion-vector information and reference
indices are derived from co-located or neighboring macroblocks.
Considering inter-view prediction based on view synthesis, an
analogous mode that derives depth and correction-vector information
from its co-located or neighboring macroblocks is considered as
well. We refer to this new coding mode as synthetic skip mode.
In the conventional skip mode as shown in FIG. 23, which applies to
both P-slices and B-slices, no residual data are coded for the
current macroblock (X) 2311. For skip in P-slices, the first entry
2304 in the reference list 2301 is chosen as the reference to
predict and derive information from, while for skip in B-slices,
the earliest entry 2305 in the reference list 2301 among
neighboring macroblocks (A, B, C) 2312-2314 is chosen as the
reference from which to predict and derive information.
Assuming that the view synthesis reference picture is not ordered
as the first entry in the reference picture list, e.g., as shown in
FIG. 11, the reference picture for skip modes in both P-slices and
B-slices would never be a view synthesis picture with existing
syntax, and following conventional decoding processes. However,
since the view synthesized picture may offer better quality
compared to a disparity-compensated picture or a motion-compensated
picture, we describe a change to the existing syntax and decoding
process to allow for skip mode based on the view synthesis
reference picture.
To utilize the skip mode with respect to a view synthesis
reference, we provide a synthetic skip mode that is signaled with
modifications to the existing mb_skip_flag. Currently, when the
existing mb_skip_flag is equal to 1, the macroblock is skipped, and
when it is equal to 0, the macroblock is not skipped.
In a first embodiment, an additional bit is added in the case when
mb_skip_flag is equal to 1 to distinguish the conventional skip
mode with the new synthetic skip mode. If the additional bit equals
1, this signals the synthetic skip mode, otherwise if the
additional bit equals 0, the conventional skip mode is used.
The above signaling scheme would work well at relatively high
bit-rates, where the number of skipped macroblocks tends to be
less. However, for lower bit-rates, the conventional skip mode is
expected to be invoked more frequently. Therefore, the signaling
scheme that includes the synthetic skip modes should not incur
additional overhead to signal the conventional skip. In a second
embodiment, an additional bit is added in the case when
mb_skip_flag is equal to 0 to distinguish the conventional non-skip
mode with the new synthetic skip mode. If the additional bit equals
1, this signals the synthetic skip mode, otherwise if the
additional bit equals 0, the conventional non-skip mode is
used.
When a high percentage of synthetic skip modes are chosen for a
slice or picture, overall coding efficiency could be improved by
reducing the overhead to signal the synthetic skip mode for each
macroblock. In a third embodiment, the synthetic skip mode is
signaled collectively for all macroblocks in a slice. This is
achieved with a slice_skip_flag included in the slice layer syntax
of the bitstream. The signaling of the slice_skip_flag would be
consistent with the mb_skip_flag described in the first and second
embodiments.
As shown in FIG. 24 when a synthetic skip mode is signaled for
P-slices, the first view synthesis reference picture 2402 in the
reference picture list 2401 is chosen as the reference instead of
the first entry 2304 in the reference picture list 2301 in the case
of conventional skip. When a synthetic skip mode is signaled for
B-slices, the earliest view synthesis reference picture 2403 in the
reference picture list 2401 is chosen as the reference instead of
the earliest entry 2305 in the reference picture list 2301 in the
case of conventional skip.
The depth and correction-vector information for synthetic skip mode
is derived as follows. A depth vector dpthLXN includes three
components (Depth, CorrX, CorrY), where Depth is a scalar value
representing the depth associated with a partition, and CorrX and
CorrY are the horizontal and vertical components of a correction
vector associated with the partition, respectively. The reference
index refIdxLX of the current partition for which the depth vector
is derived is assigned as the first or earliest view synthesis
reference picture depending on whether the current slices is a
P-slice or B-slice, as described above.
Inputs to this process are the neighboring partitions A 2312, B
2313 and C 2314, the depth vector for each of the neighboring
partitions denoted dpthLXN (with N being replaced by A, B, or C),
the reference indices refIdxLXN (with N being replaced by A, B, or
C) of the neighboring partitions, and the reference index refIdxLX
of the current partition.
Output of this process is the depth vector prediction dpthpLX. The
variable dpthpLX is derived as follows. When neither the B 2313 nor
C 2314 neighboring partition is available and the A 2312
neighboring partition is available, then the following assignments
apply: dpthLXB=dpthLXA and dpthLXC=depthLXA, refIdxLXB=refIdxLA and
refIdxLXC=refIdxLA.
If refIdxLXN is a reference index of a multiview reference picture
from which a synthetic multiview reference picture with the
reference index refIdxLX is to be synthesized, refIdxLXN is
considered equal to refIdxLX and its associated depth vector
dpthLXN is derived by converting the disparity from the reference
picture with the reference index refIdxLXN to an equivalent depth
vector associated with the reference picture with the reference
index refIdxLX.
Depending on reference indices refIdxLXA, refIdxLXB, or refIdxLXC,
the following applies. If one and only one of the reference indices
refIdxLXA, refIdxLXB, or refIdxLXC is equal to the reference index
refIdxLX of the current partition, the following applies. Let
refIdxLXN be the reference index that is equal to refIdxLX, the
depth vector dpthLXN is assigned to the depth vector prediction
dpthpLX. Otherwise, each component of the depth vector prediction
dpthpLX is given by the median of the corresponding vector
components of the depth vector dpthLXA, dpthLXB, and dpthLXC.
Direct Mode
Similar to skip mode, the conventional direct mode for B-slices
also derives motion-vector information and reference indices from
neighboring macroblocks. Direct mode differs from skip mode in that
residual data are also present. For the same reasons that we
provide the synthetic skip mode, we also describe an analogous
extension of direct mode that we refer to as synthetic direct
mode.
To invoke the conventional direct mode, a macroblock is coded as
non-skip. Direct mode could then be applied to both 16.times.16
macroblocks and 8.times.8 blocks. Both of these direct modes are
signaled as a macroblock mode.
In a first embodiment, the method for signaling the synthetic
direct mode is done by adding an additional mode to the list of
candidate macroblock modes.
In a second embodiment, the method for signaling the synthetic
direct mode is done by signaling an additional flag indicating that
the 16.times.16 macroblock or 8.times.8 block is coded as synthetic
direct mode.
When a synthetic direct mode is signaled for B-slices, the earliest
view synthesis reference pictures in the reference picture lists
are chosen as the references instead of the earliest entries in the
reference picture lists in the case of conventional direct
mode.
The same processes for deriving the depth and correction vector
information as synthetic skip mode is done for synthetic direct
modes.
Although the invention has been described by way of examples of
preferred embodiments, it is to be understood that various other
adaptations and modifications can be made within the spirit and
scope of the invention. Therefore, it is the object of the appended
claims to cover all such variations and modifications as come
within the true spirit and scope of the invention.
TABLE-US-00001 Appendix A Macroblock Syntax and Semantics C
Descriptor mb_pred( mb_type ) { . . . } else if( MbPartPredMode(
mb_type, 0 ) != Direct ) { . . . for( mbPartIdx = 0; mbPartIdx <
NumMbPart( mb_type ); mbPartIdx++) if( MbPartPredMode ( mb_type,
mbPartIdx ) != Pred_L1 ) { if
(multiview_type(view_id(ref_idx_l0[mbPartIdx]) )== 1) {
depthd_l0[mbPartIdx][0] 2 se(v) | ae(v) for( compIdx = 0; compIdx
< 2; compIdx++) corr_vd_10[mbPartIdx][0][compIdx] 2 se(v) |
ae(v) } else for( compIdx = 0; compIdx < 2; compIdx++) mvd_l0[
mbPartIdx ][ 0 ] [ compIdx ] 2 se(v) | ae(v) } for( mbPartIdx = 0;
mbPartIdx < NumMbPart( mb_type ); mbPartIdx++) if(
MbPartPredMode( mb_type, mbPartIdx ) != Pred_L0 { If
(multiview_type (view_id(ref_idx_l1[mbPartIdx]) )== 1) {
depthd_l1[mbPartIdx][0] 2 se(v) | ae(v) for( compIdx = 0; compIdx
< 2; compIdx++) corr_vd_l1[mbPartIdx][0][compIdx] 2 se(v) |
ae(v) } else for( compIdx = 0; compIdx < 2; compIdx++) mvd_l1[
mbPartIdx ][ 0 ][ compIdx ] 2 se(v) | ae(v) } } } sub_mb_pred(
mb_type ) { . . . for( mbPartIdx = 0; mbPartIdx < 4;
mbPartIdx++) if( sub_mb_type[ mbPartIdx ] != B_Direct_8x8
&& SubMbPredMode( sub_mb_type[ mbPartIdx ] ) != Pred_L1 )
for( subMbPartIdx = 0; subMbPartIdx < NumSubMbPart( sub_mb_type[
mbPartIdx ] ); subMbPartIdx++) { if ( multiview_type
(view_id(ref_idx_l0[mbPartIdx]) )== 1) {
depthd_l0[mbPartIdx][subMbPartIdx] 2 se(v) | ae(v) for( compIdx =
0; compIdx < 2; compIdx++)
corr_vd_l0[mbPartIdx][subMbPartIdx][compIdx] 2 se(v) | ae(v) } else
for( compIdx = 0; compIdx < 2; compIdx++) mvd_l0[ mbPartId ][
subMbPartIdx ][ compIdx ] 2 se(v) | ae(v) } for( mbPartIdx = 0;
mbPartIdx < 4; mbPartIdx++ ) if( sub_mb_type[ mbPartIdx ] !=
B_Direct_8x8 && SubMbPredMode( sub_mb_type[ mbPartIdx ] )
!= Pred_L0 ) for( subMbPartIdx = 0: subMbPartIdx < NumSubMbPart(
sub_mb_type[ mbPartIdx ] ); subMbPartIdx++) { if (multiview_type
(view_id(ref_idx_l1[mbPartIdx]) )==1) {
depthd_l1[mbPartIdx][subMbPartIdx] 2 se(v) | ae(v) for( compIdx =
0; compIdx < 2; compIdx++)
corr_vd_l1[mbPartIdx][subMbPartIdx][compIdx] 2 se(v) | ae(v) } else
for( compIdx = 0; compIdx < 2; compIdx++) mvd_l1[ mbPartIdx ][
subMbPartIdx ][ compIdx ] 2 se(v) | ae(v) }
Additions to Subclause 7.4.5.1 Macroblock Prediction Semantics:
depthd_l0[mbPartIdx][0] specifies the difference between a depth
value to be used and its prediction. The index mbPartIdx specifies
to which macroblock partition depthd_l0 is assigned. The
petitioning of the macroblock is specified by mb_type.
depthd_l1[mbPartIdx][0] has the same semantics as depthd_l0, with
l0 replaced by l1. corr_vd_l0[mbPartIdx][0][compIdx] specifies the
difference between a correction-vector component to be used and its
prediction. The index mbPartIdx specifies to which macroblock
partition corr_vd_l0 is assigned. The partitioning of the
macroblock is specified by mb_type. The horizontal correction
vector component difference is decoded first in decoding order and
is assigned CompIdx=0. The vertical correction vector component is
decoded second in decoding order and is assigned CompIdx=1.
corr_vd_l1[mbPartIdx][0][compIdx] has the same semantics as
corr_vd_l0, with l0 replaced by l1. Additions to Subclause 7.4.5.2
Sub-Macro Block Prediction Semantics:
depthd_l0[mbPartIdx][subMbPartIdx] has the same semantics as
depthd_l0, except that it is applied to the sub-macroblock
partition index with subMbPartIdx. The indices mbPartIdx and
subMbPartIdx specify to which macroblock partition and
sub-macroblock partition depthd_l0 is assigned. depthd_l1
[mbPartIdx][subMbPartIdx] has the same semantics as depthd.sub.--0
with l0 replaced by l1.
corr_vd_l0[mbPartIdx][subMbPartIdx][compIdx] has the same semantics
as corr_vd_l0, except that it is applied to the sub-macroblock
partition index with subMbPartIdx. The indices mbPartIdx and
subMbPartIdx specify to which macroblock partition and
sub-macroblock partition corr_vd.sub.--0 is assigned.
corr_vd_l1[mbPartIdx][subMbPartIdx][compIdx] has the same semantics
as corr_vd_l1, with l0 replaced by l1. View Parameter Set
Syntax:
TABLE-US-00002 Descriptor view_parameter_set( ) { view_id ue(v)
multiview_type u (l) if (multiview_type) { multiview_synth_ref0
ue(v) multiview_synth_ref1 ue(v) } num_multiview_refs_for_list0
ue(v) num_multiview_refs_for_list1 ue(v) for( i = 0; i <
num_multiview_refs_for_list0; i++ ) { reference_view_for_list_0[i]
ue(v) } for( i = 0; i < num_multiview_refs_for_list1; i++ ) {
reference_view_for_list_1[i] ue(v) } }
Additions to View Parameter Set Semantics: multiview_type equal to
1 specifies that the current view is synthesized from other views.
multiview_type equal to 0 specifies that the current view is not a
synthesized one. multiview_synth_ref0 specifies an index for the
first view to be used for synthesis. multiview_synth_ref1 specifies
an index for the second view to be used for synthesis.
* * * * *